| |
| |
| """ |
| mysql_006 ground truth: 推理服务流量与资源预测-P90分位数与节假日降级 |
| |
| Task: |
| 以20260507为预测基准日，为所有在线推理服务生成未来14天的流量与资源使用量预测数据。 |
| 预测采用断点检测（CPD）结果之后的历史数据，区分节假日/工作日类型计算P90分位数， |
| 当某类型历史数据不足时降级使用另一类型数据，输出10分钟和小时两种时间粒度的预测结果。 |
| """ |
| import pymysql |
| import sys |
|
|
| DB_NAME = "internal_platform_db" |
| INPUT_TABLE_FEATURE = "dwm_gputj_platform_gpu_base_feature_agg_v2_mysql_006" |
| INPUT_TABLE_SERVICE = "dwd_aide_inferencev2_done_service_info_h_mysql_006" |
| INPUT_TABLE_CPD = "dwd_gputj_platform_metric_cpd_offline_v2_mysql_006" |
| INPUT_TABLE_HOLIDAY = "dim_holiday_list_mysql_006" |
| OUTPUT_TABLE = "dwm_gputj_platform_model_prediction_long_cpd_mysql_006" |
|
|
| MYSQL_CONFIG = { |
| "host": "localhost", |
| "port": 3306, |
| "user": "root", |
| "password": "root123", |
| "charset": "utf8mb4", |
| } |
|
|
| gt_sql = f""" |
| INSERT INTO {DB_NAME}.{OUTPUT_TABLE} |
| (instance_uuid, service_name, workload_name, namespace, agg_time, agg_type, |
| is_holiday, day_of_week, prediction_type, nv_inference_count_model_avg_p90, |
| statistic_time_count, nv_inference_request_duration_ms_model_avg, |
| nv_inference_queue_duration_ms_model_avg, num_queued_reqs_model_avg, |
| nv_inference_request_success_model_avg, nv_inference_request_failure_model_avg, |
| nv_inference_request_duration_ms_perreq_avg, nv_inference_queue_duration_ms_perreq_avg, |
| nv_inference_request_duration_ms_perreq_p95, nv_inference_queue_duration_ms_perreq_p95, |
| nv_inference_request_success_model_max, nv_inference_request_failure_model_max, |
| DCGM_FI_DEV_GPU_UTIL_pod_avg, k8s_container_bs_rate_cpu_core_used_request_pod_avg, |
| k8s_container_rate_mem_working_set_request_pod_avg, k8s_dcgm_fi_dev_fb_util_pod_avg, |
| k8s_container_vgpu_gpu_util_pod_avg, cpd_agg_time, feature_total, feature_holiday_total, dt) |
| WITH RECURSIVE offsets AS ( |
| SELECT 0 AS n |
| UNION ALL SELECT n + 1 FROM offsets WHERE n < 13 |
| ), |
| future_dates AS ( |
| SELECT DATE_ADD(STR_TO_DATE('20260507', '%Y%m%d'), INTERVAL n DAY) AS target_date |
| FROM offsets |
| ), |
| time_template AS ( |
| SELECT SUBSTRING_INDEX(agg_time, ' ', -1) AS time_part, agg_type |
| FROM {DB_NAME}.{INPUT_TABLE_FEATURE} |
| WHERE dt >= '20260502' AND dt < '20260507' AND agg_time > '' |
| GROUP BY SUBSTRING_INDEX(agg_time, ' ', -1), agg_type |
| ), |
| time_list AS ( |
| SELECT CONCAT(fd.target_date, ' ', tt.time_part) AS agg_time, tt.agg_type, fd.target_date |
| FROM future_dates fd |
| CROSS JOIN time_template tt |
| ), |
| service_list AS ( |
| SELECT DISTINCT name AS service_name |
| FROM {DB_NAME}.{INPUT_TABLE_SERVICE} |
| WHERE dt = '2026050723' AND name <> '' |
| ), |
| cpd AS ( |
| SELECT service_name, agg_time |
| FROM ( |
| SELECT service_name, agg_time, |
| ROW_NUMBER() OVER (PARTITION BY service_name ORDER BY dt DESC, agg_time DESC) AS r1 |
| FROM ( |
| SELECT service_name, dt, agg_time, |
| ROW_NUMBER() OVER (PARTITION BY service_name, dt, metric_name ORDER BY metric_value ASC) AS r |
| FROM {DB_NAME}.{INPUT_TABLE_CPD} |
| WHERE dt <= '20260507' AND dt >= '20260421' |
| AND msg = 'success' |
| AND agg_time < '20260507' |
| ) t0 |
| WHERE r = 1 |
| ) t |
| WHERE r1 = 1 |
| ), |
| feature AS ( |
| SELECT t.service_name, instance_uuid, trial_job_name AS workload_name, namespace, |
| t.agg_time, agg_type, nv_inference_count_model_avg, |
| nv_inference_request_duration_ms_model_avg, nv_inference_queue_duration_ms_model_avg, |
| num_queued_reqs_model_avg, nv_inference_request_success_model_avg, |
| nv_inference_request_failure_model_avg, nv_inference_request_duration_ms_perreq_avg, |
| nv_inference_queue_duration_ms_perreq_avg, nv_inference_request_duration_ms_perreq_p95, |
| nv_inference_queue_duration_ms_perreq_p95, nv_inference_request_success_model_max, |
| nv_inference_request_failure_model_max, dcgm_fi_dev_gpu_util_pod_avg, |
| k8s_container_bs_rate_cpu_core_used_request_pod_avg, |
| k8s_container_rate_mem_working_set_request_pod_avg, |
| k8s_dcgm_fi_dev_fb_util_pod_avg, k8s_container_vgpu_gpu_util_pod_avg |
| FROM {DB_NAME}.{INPUT_TABLE_FEATURE} t |
| JOIN cpd ON cpd.service_name = t.service_name AND cpd.agg_time <= t.agg_time |
| WHERE dt >= '20260421' AND dt <= '20260507' |
| AND nv_inference_count_model_avg >= 0 |
| AND agg_type IN (1, 2) |
| ), |
| base AS ( |
| SELECT '20260507' AS dt, |
| instance_uuid, service_name, workload_name, namespace, agg_time, agg_type, |
| today_holiday_date, |
| MOD(DATEDIFF(SUBSTRING_INDEX(agg_time, ' ', 1), '2019-12-30'), 7) + 1 AS day_of_week, |
| 'request_model_count' AS prediction_type, |
| nv_inference_count_model_avg, |
| CASE |
| WHEN today_holiday_date = 1 AND feature_holiday_total > 0 THEN feature_holiday_total |
| WHEN today_holiday_date = 1 AND feature_holiday_total = 0 THEN feature_weekday_total |
| WHEN today_holiday_date = 0 AND feature_weekday_total > 0 THEN feature_weekday_total |
| WHEN today_holiday_date = 0 AND feature_weekday_total = 0 THEN feature_holiday_total |
| END AS statistic_time_count, |
| nv_inference_request_duration_ms_model_avg, nv_inference_queue_duration_ms_model_avg, |
| num_queued_reqs_model_avg, nv_inference_request_success_model_avg, |
| nv_inference_request_failure_model_avg, nv_inference_request_duration_ms_perreq_avg, |
| nv_inference_queue_duration_ms_perreq_avg, nv_inference_request_duration_ms_perreq_p95, |
| nv_inference_queue_duration_ms_perreq_p95, nv_inference_request_success_model_max, |
| nv_inference_request_failure_model_max, dcgm_fi_dev_gpu_util_pod_avg, |
| k8s_container_bs_rate_cpu_core_used_request_pod_avg, |
| k8s_container_rate_mem_working_set_request_pod_avg, |
| k8s_dcgm_fi_dev_fb_util_pod_avg, k8s_container_vgpu_gpu_util_pod_avg, |
| total AS feature_total, feature_holiday_total |
| FROM ( |
| SELECT *, |
| total - feature_holiday_total AS feature_weekday_total, |
| CEIL((total - feature_holiday_total) * 0.9) AS feature_weekday_index, |
| CEIL(feature_holiday_total * 0.9) + (total - feature_holiday_total) AS feature_holiday_index |
| FROM ( |
| SELECT *, |
| ROW_NUMBER() OVER (PARTITION BY service_name, agg_time, agg_type |
| ORDER BY feature_holiday_date ASC, nv_inference_count_model_avg ASC) AS r, |
| COUNT(*) OVER (PARTITION BY service_name, agg_time, agg_type) AS total, |
| SUM(feature_holiday_date) OVER (PARTITION BY service_name, agg_time, agg_type) AS feature_holiday_total |
| FROM ( |
| SELECT time_list.agg_time, feature.agg_type, |
| MAX(feature.instance_uuid) AS instance_uuid, |
| feature.service_name, |
| MAX(feature.workload_name) AS workload_name, |
| MAX(feature.namespace) AS namespace, |
| feature.agg_time AS feature_agg_time, |
| MAX(CASE WHEN holiday_today.holiday_date > '' THEN 1 ELSE 0 END) AS today_holiday_date, |
| MAX(feature.nv_inference_count_model_avg) AS nv_inference_count_model_avg, |
| MAX(feature.nv_inference_request_duration_ms_model_avg) AS nv_inference_request_duration_ms_model_avg, |
| MAX(feature.nv_inference_queue_duration_ms_model_avg) AS nv_inference_queue_duration_ms_model_avg, |
| MAX(feature.num_queued_reqs_model_avg) AS num_queued_reqs_model_avg, |
| MAX(feature.nv_inference_request_success_model_avg) AS nv_inference_request_success_model_avg, |
| MAX(feature.nv_inference_request_failure_model_avg) AS nv_inference_request_failure_model_avg, |
| MAX(feature.nv_inference_request_duration_ms_perreq_avg) AS nv_inference_request_duration_ms_perreq_avg, |
| MAX(feature.nv_inference_queue_duration_ms_perreq_avg) AS nv_inference_queue_duration_ms_perreq_avg, |
| MAX(feature.nv_inference_request_duration_ms_perreq_p95) AS nv_inference_request_duration_ms_perreq_p95, |
| MAX(feature.nv_inference_queue_duration_ms_perreq_p95) AS nv_inference_queue_duration_ms_perreq_p95, |
| MAX(feature.nv_inference_request_success_model_max) AS nv_inference_request_success_model_max, |
| MAX(feature.nv_inference_request_failure_model_max) AS nv_inference_request_failure_model_max, |
| MAX(feature.dcgm_fi_dev_gpu_util_pod_avg) AS dcgm_fi_dev_gpu_util_pod_avg, |
| MAX(feature.k8s_container_bs_rate_cpu_core_used_request_pod_avg) AS k8s_container_bs_rate_cpu_core_used_request_pod_avg, |
| MAX(feature.k8s_container_rate_mem_working_set_request_pod_avg) AS k8s_container_rate_mem_working_set_request_pod_avg, |
| MAX(feature.k8s_dcgm_fi_dev_fb_util_pod_avg) AS k8s_dcgm_fi_dev_fb_util_pod_avg, |
| MAX(feature.k8s_container_vgpu_gpu_util_pod_avg) AS k8s_container_vgpu_gpu_util_pod_avg, |
| MAX(CASE WHEN holiday_feature.holiday_date > '' THEN 1 ELSE 0 END) AS feature_holiday_date |
| FROM service_list |
| CROSS JOIN time_list |
| JOIN feature ON service_list.service_name = feature.service_name |
| AND time_list.agg_type = feature.agg_type |
| AND SUBSTRING_INDEX(time_list.agg_time, ' ', -1) = SUBSTRING_INDEX(feature.agg_time, ' ', -1) |
| LEFT JOIN {DB_NAME}.{INPUT_TABLE_HOLIDAY} holiday_today |
| ON SUBSTRING_INDEX(time_list.agg_time, ' ', 1) = holiday_today.holiday_date |
| LEFT JOIN {DB_NAME}.{INPUT_TABLE_HOLIDAY} holiday_feature |
| ON SUBSTRING_INDEX(feature.agg_time, ' ', 1) = holiday_feature.holiday_date |
| GROUP BY time_list.agg_time, feature.agg_time, feature.agg_type, feature.service_name |
| ) t1 |
| ) t2 |
| ) t3 |
| WHERE r = CASE |
| WHEN today_holiday_date = 1 AND feature_holiday_total >= 1 THEN feature_holiday_index |
| WHEN today_holiday_date = 1 AND feature_holiday_total < 1 THEN feature_weekday_index |
| WHEN today_holiday_date = 0 AND feature_weekday_total >= 1 THEN feature_weekday_index |
| WHEN today_holiday_date = 0 AND feature_weekday_total < 1 THEN feature_holiday_index |
| END |
| ) |
| SELECT instance_uuid, base.service_name, base.workload_name, base.namespace, |
| base.agg_time, agg_type, today_holiday_date, day_of_week, prediction_type, |
| nv_inference_count_model_avg, statistic_time_count, |
| nv_inference_request_duration_ms_model_avg, nv_inference_queue_duration_ms_model_avg, |
| num_queued_reqs_model_avg, nv_inference_request_success_model_avg, |
| nv_inference_request_failure_model_avg, nv_inference_request_duration_ms_perreq_avg, |
| nv_inference_queue_duration_ms_perreq_avg, nv_inference_request_duration_ms_perreq_p95, |
| nv_inference_queue_duration_ms_perreq_p95, nv_inference_request_success_model_max, |
| nv_inference_request_failure_model_max, dcgm_fi_dev_gpu_util_pod_avg, |
| k8s_container_bs_rate_cpu_core_used_request_pod_avg, |
| k8s_container_rate_mem_working_set_request_pod_avg, |
| k8s_dcgm_fi_dev_fb_util_pod_avg, k8s_container_vgpu_gpu_util_pod_avg, |
| cpd.agg_time AS cpd_agg_time, feature_total, feature_holiday_total, base.dt |
| FROM base |
| LEFT JOIN cpd ON base.service_name = cpd.service_name |
| """ |
|
|
|
|
| def main(): |
| conn = pymysql.connect(**MYSQL_CONFIG) |
|
|
| try: |
| with conn.cursor() as cur: |
| |
| cur.execute(f""" |
| CREATE TABLE IF NOT EXISTS {DB_NAME}.{OUTPUT_TABLE} ( |
| instance_uuid VARCHAR(256), |
| service_name VARCHAR(256), |
| workload_name VARCHAR(256), |
| namespace VARCHAR(256), |
| agg_time VARCHAR(256), |
| agg_type BIGINT, |
| is_holiday BIGINT, |
| day_of_week BIGINT, |
| prediction_type VARCHAR(256), |
| nv_inference_count_model_avg_p90 DOUBLE, |
| statistic_time_count BIGINT, |
| nv_inference_request_duration_ms_model_avg DOUBLE, |
| nv_inference_queue_duration_ms_model_avg DOUBLE, |
| num_queued_reqs_model_avg DOUBLE, |
| nv_inference_request_success_model_avg DOUBLE, |
| nv_inference_request_failure_model_avg DOUBLE, |
| nv_inference_request_duration_ms_perreq_avg DOUBLE, |
| nv_inference_queue_duration_ms_perreq_avg DOUBLE, |
| nv_inference_request_duration_ms_perreq_p95 DOUBLE, |
| nv_inference_queue_duration_ms_perreq_p95 DOUBLE, |
| nv_inference_request_success_model_max DOUBLE, |
| nv_inference_request_failure_model_max DOUBLE, |
| DCGM_FI_DEV_GPU_UTIL_pod_avg DOUBLE, |
| k8s_container_bs_rate_cpu_core_used_request_pod_avg DOUBLE, |
| k8s_container_rate_mem_working_set_request_pod_avg DOUBLE, |
| k8s_dcgm_fi_dev_fb_util_pod_avg DOUBLE, |
| k8s_container_vgpu_gpu_util_pod_avg DOUBLE, |
| cpd_agg_time VARCHAR(256), |
| feature_total BIGINT, |
| feature_holiday_total BIGINT, |
| dt VARCHAR(256) |
| ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 |
| """) |
|
|
| |
| cur.execute(f"TRUNCATE TABLE {DB_NAME}.{OUTPUT_TABLE}") |
| cur.execute(gt_sql) |
|
|
| conn.commit() |
|
|
| |
| with conn.cursor() as cur: |
| cur.execute(f"SELECT COUNT(*) FROM {DB_NAME}.{OUTPUT_TABLE}") |
| count = cur.fetchone()[0] |
| print(f"mysql_006 ground_truth done: {count} rows written to output table") |
| except Exception as e: |
| print(f"ground_truth error: {e}", file=sys.stderr) |
| conn.rollback() |
| sys.exit(1) |
| finally: |
| conn.close() |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
